Incorrect version number given to sync_pgdata() in pg_combinebackup.c
Hi all,
While hacking on a different thing that touched pg_combinebackup, I
have bumped into a silly bug.
To keep it short, the version number is calculated based on this code
in read_pg_version_file(), where "version" is the result of strtoul()
applied to the contents of PG_VERSION:
return version * 10000;
For v18, this would return 180000, which is fine.
A bit later on, we do that, which is not fine:
sync_pgdata(opt.output, version * 10000, opt.sync_method, true);
This leads to a version number higher than expected, multiplied twice.
This was harmless, because sync_pgdata uses the version number to make
the difference between pg_wal/ and pg_xlog/, and pg_combinebackup does
not support versions older than v10, which is exactly where the
renamed happened. Hence, even if the version number was too high, we
always expect to flush pg_wal/.
Trivial patch attached, for a backpatch down to where pg_combinebackup
has been introduced.
Thoughts?
--
Michael
Attachments:
pg_combinebackup-version.patchtext/x-diff; charset=us-asciiDownload+1-1
On Oct 10, 2025, at 14:21, Michael Paquier <michael@paquier.xyz> wrote:
Hi all,
While hacking on a different thing that touched pg_combinebackup, I
have bumped into a silly bug.To keep it short, the version number is calculated based on this code
in read_pg_version_file(), where "version" is the result of strtoul()
applied to the contents of PG_VERSION:
return version * 10000;For v18, this would return 180000, which is fine.
A bit later on, we do that, which is not fine:
sync_pgdata(opt.output, version * 10000, opt.sync_method, true);This leads to a version number higher than expected, multiplied twice.
This was harmless, because sync_pgdata uses the version number to make
the difference between pg_wal/ and pg_xlog/, and pg_combinebackup does
not support versions older than v10, which is exactly where the
renamed happened. Hence, even if the version number was too high, we
always expect to flush pg_wal/.Trivial patch attached, for a backpatch down to where pg_combinebackup
has been introduced.Thoughts?
--
Michael
<pg_combinebackup-version.patch>
Yeah, looks like a stupid bug. read_pg_version_file() has multiplied 10000 to version number.
Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/